By Greg Bensinger SAN FRANCISCO, Sept 3 (Reuters) - OpenAI on Thursday unveiled a new artificial intelligence model it calls its best yet but cautioned that it also sometimes attempts to evade human monitoring,

The numbers on OpenAI's blog about Astra kept changing. In some cases, the changes made Astra look better and rival models look worse.

Safety researchers say OpenAI's Astra reasons less in visible text. Its chief scientist co-authored the 2025 paper asking developers to report this.

OpenAI says access has reached Plus and Business subscribers, but the new model is not available in every ChatGPT mode.

GPT-6 Astra leads every comparable benchmark against Claude Fable 5.1, Claude Fable 5, and Claude Opus 5.

OpenAI launched GPT-6 Astra on Sept. 3 with record benchmarks, a 'Critical' cyber rating, and Brockman's declaration of an AGI era. Access stays limited.

OpenAI launches ChatGPT 6 Astra for complex professional workflows. The new AI model reduces task completion time by 47 percent for users.

OpenAI launched GPT-6 Astra with advanced computer-use capabilities as Greg Brockman declared the “AGI era” has arrived, though major questions remain.

OpenAI launches GPT-6 Astra to ChatGPT subscribers, scoring 100% on ExploitBench and completing tasks 47% faster than GPT-5.6 Sol amid early

OpenAI's GPT-6 Astra operates software like a human — spawning sub-agents to compile code and run test suites on local CPUs.